UCLA-VAIL/SidewalkBench-GS-Assets

SidewalkBench GS Assets

The real-world scanned scenes of SidewalkBench's 3D Gaussian-splat evaluation (sidewalkbench/gs): 11 sidewalk scenes with 30 short-horizon routes (10 straight, 10 curved, 10 crosswalk) and 5 long-horizon routes, the pedestrian avatars and the traffic cone (~12 GB).

Download

From the SidewalkBench repository root:

uv run hf download UCLA-VAIL/SidewalkBench-GS-Assets --repo-type dataset --local-dir gs_assets

The evaluation reads gs_assets/ (see the SidewalkBench README), e.g.

sidewalkbench/gs/run_gs_eval.sh FlowPilot ucla_crosswalk --save_video                  # short horizon
sidewalkbench/gs/run_gs_eval.sh FlowPilot dark --mode long_horizon --save_video        # long horizon

Layout

<scene>/
  gs.ply                   3D Gaussian splat (the robot's camera observations)
  mesh.ply                 collision mesh (physics, ground height)
  semantic_grid_0p1m.npz   semantic grid, 0.1 m (walkable = sidewalk + crosswalk): the off-sidewalk check
  walkable.npz             geometric walkability grid, 0.05 m: the pedestrians' motion
  centerlines.json         the scene's sidewalk centerlines
  annotation.json          routes (start, goal, centerline, region, pedestrians) and cones
pedestrians/<identity>/    baked HUGS avatars: walk.npz (walk cycle), stand.npz (standing pose)
cone/                      traffic cone: traffic_cone.glb (mesh) and traffic_cone_splat.ply (splat)

Scenes

Scene	Short-horizon routes	Long-horizon routes	Cones	Pedestrians	Gaussians	Size
9am	2	0	1	0	22.0M	1.36 GB
beverly_hills_crosswalk	3	0	7	0	11.9M	0.76 GB
beverly_hills_intersection	4	0	0	0	14.7M	0.93 GB
cars	4	0	1	0	12.2M	0.76 GB
dark	2	1	21	3	28.8M	1.85 GB
hill	2	0	8	0	25.7M	1.64 GB
loop	4	0	4	0	10.3M	0.64 GB
midvale	1	0	4	0	19.8M	1.24 GB
triangle	3	0	5	0	13.2M	0.84 GB
ucla_crosswalk	5	0	0	0	5.9M	0.38 GB
ucla_sidewalk_backward	0	4	7	19	10.0M	0.64 GB

Pedestrian identities: bike, citron, lab, parkinglot, seattle (0.99 GB).

License

MIT.

Citation

To be added.